Cross-Directional Aggregation and Feature-Loss Optimization for Remote Sensing Semantic Segmentation

The complexity of object categories in high-resolution remote sensing frequently gives rise to discontinuous boundaries in segmentation results. Furthermore, the inability to model global contextual morphological continuities along strip objects is a primary cause of segmentation fractures. Therefore, we propose a remote sensing semantic segmentation method with cross-directional aggregation and feature-loss optimization. First, we design a two-stage cross-directional strip feature extraction module (TCSFE), which is dedicated to modeling the global contextual information of strip objects. Specifically, our proposed module employs cross-directional attention interactions to capture global dependencies along both the horizontal and vertical directions. Second, we construct a hybrid attention-guided semantic injection module (HAGSI), which employs a dual-branch collaborative mechanism that integrates hybrid attention to enhance long-range feature modeling. Finally, we propose a feature-loss dual optimization boosted superpixel edge module (FDOSE). It employs feature fusion optimization and unified superpixel loss optimization to jointly optimize the final segmentation result. Experimental results on the Potsdam and Vaihingen datasets and the self-built TSRSD (Taiyuan Satellite Remote Sensing Dataset) demonstrate that our proposed method achieves state-of-the-art (SOTA) performance on the mIoU and Recall metrics, with mIoU scores of 79.93%, 80.25%, and 64.89% on Potsdam, Vaihingen, and TSRSD, respectively, and Recall scores of 87.70%, 88.94%, and 73.25%, respectively. Moreover, it exhibits an advantage in boundary continuity.